AI Generated Pokemon Sprites with GPT-2 (matthew.rayfield.world)

Indie developer Matthew Rayfield repurposed GPT-2 to generate Pokemon-style sprites by converting 32x32 pixel images into plain-text sequences, training the language model on those textified sprites, and converting the outputs back into images. The line format encodes row number, orientation, and one-character pixel tokens (e.g., ~, >, !, `), and his generation pipeline seeds a few lines, filters “good” outputs, has GPT-2 continue the sequence, and repeats until all 64 rows are produced. After hours of Colab training he produced thousands of quirky sprites; while most are noisy, many capture the visual “essence” of Pokemon and a hand-picked subset was even redrawn by an artist into polished illustrations. Technically this is a neat demonstration that autoregressive text transformers can model low-resolution image structure when images are serialized as token sequences—no bespoke vision model required. It highlights practical points: careful tokenization/format matters, output normalization and filtering significantly improve results, and small-image regimes are tractable on consumer GPUs. Limitations include noisy outputs and selective showcasing of good examples, but the project (code and Colab notebooks) is open-sourced, offering an accessible playground for exploratory multimodal work and pointing toward more advanced approaches like OpenAI’s ImageGPT.